Re: [PATCH 1/6] hash.h: remove unused define directive

[David Gow]

On Thu, Aug 26, 2021 at 9:26 AM Isabella Basso <isabellabdoamaral@xxxxxx> wrote:
>
> The HAVE_ARCH_HASH_32 (single underscore) define hasn't been used for
> any known supported architectures that have their own hash function
> implementation (i.e. m68k, Microblaze, H8/300, pa-risc) since George's
> patch [1], which introduced it.
>
> The supported 32-bit architectures from the list above have only been
> making use of the (more general) HAVE_ARCH__HASH_32 define, which only
> lacks the right shift operator, that wasn't targeted for optimizations
> so far.
>
> [1] https://lore.kernel.org/lkml/20160525073311.5600.qmail@xxxxxxxxxxxxxxxxxxxxxx/
>
> Co-developed-by: Augusto Durães Camargo <augusto.duraes33@xxxxxxxxx>
> Signed-off-by: Augusto Durães Camargo <augusto.duraes33@xxxxxxxxx>
> Co-developed-by: Enzo Ferreira <ferreiraenzoa@xxxxxxxxx>
> Signed-off-by: Enzo Ferreira <ferreiraenzoa@xxxxxxxxx>
> Signed-off-by: Isabella Basso <isabellabdoamaral@xxxxxx>
> ---

I'm not familiar with the hash functions here, so take this with the
appropriate heap of salt, but it took me a little while to understand
exactly what this is doing.

As I understand it:
- There are separate __hash_32() and hash_32() functions.
- Both of these have generic implementations, which can optionally be
overridden by an architecture-specific optimised version.
- There aren't any architectures which provide an optimised hash_32()
implementation.
- This patch therefore removes support for architecture-specific
hash_32() implementations, and leaves only the generic implementation.
- This generic implementation of hash_32() itself relies on
__hash_32(), which may still be optimised.

Could the commit description be updated to make this a bit clearer?
While we are getting rid of the HAVE_ARCH_HASH_32 #define, that seems
to be a side-effect/implementation detail of removing support for
architecture-specific hash_32() implementations...

The other wild, out-there option would be to remove __hash_32()
entirely and make everything use hash_32(), which then could have
architecture-specific implementations. A quick grep reveals that
there's only one use of __hash_32() outside of the hashing code itself
(in fs/namei.c). This would be much more consistent with what
hash_64() does, but also would be significantly more work, and
potentially could have some implication (full_name_hash() performance
maybe?) which I'm not aware of. So it's possibly not worth it.

Cheers,
-- David

>  include/linux/hash.h       |  5 +----
>  lib/test_hash.c            | 24 +-----------------------
>  tools/include/linux/hash.h |  5 +----
>  3 files changed, 3 insertions(+), 31 deletions(-)
>
> diff --git a/include/linux/hash.h b/include/linux/hash.h
> index ad6fa21d977b..38edaa08f862 100644
> --- a/include/linux/hash.h
> +++ b/include/linux/hash.h
> @@ -62,10 +62,7 @@ static inline u32 __hash_32_generic(u32 val)
>         return val * GOLDEN_RATIO_32;
>  }
>
> -#ifndef HAVE_ARCH_HASH_32
> -#define hash_32 hash_32_generic
> -#endif
> -static inline u32 hash_32_generic(u32 val, unsigned int bits)
> +static inline u32 hash_32(u32 val, unsigned int bits)
>  {
>         /* High bits are more random, so use them. */
>         return __hash_32(val) >> (32 - bits);
> diff --git a/lib/test_hash.c b/lib/test_hash.c
> index 0ee40b4a56dd..d4b0cfdb0377 100644
> --- a/lib/test_hash.c
> +++ b/lib/test_hash.c
> @@ -94,22 +94,7 @@ test_int_hash(unsigned long long h64, u32 hash_or[2][33])
>                         pr_err("hash_32(%#x, %d) = %#x > %#x", h0, k, h1, m);
>                         return false;
>                 }
> -#ifdef HAVE_ARCH_HASH_32
> -               h2 = hash_32_generic(h0, k);
> -#if HAVE_ARCH_HASH_32 == 1
> -               if (h1 != h2) {
> -                       pr_err("hash_32(%#x, %d) = %#x != hash_32_generic() "
> -                               " = %#x", h0, k, h1, h2);
> -                       return false;
> -               }
> -#else
> -               if (h2 > m) {
> -                       pr_err("hash_32_generic(%#x, %d) = %#x > %#x",
> -                               h0, k, h1, m);
> -                       return false;
> -               }
> -#endif
> -#endif
> +
>                 /* Test hash_64 */
>                 hash_or[1][k] |= h1 = hash_64(h64, k);
>                 if (h1 > m) {
> @@ -227,13 +212,6 @@ test_hash_init(void)
>  #else
>         pr_info("__hash_32() has no arch implementation to test.");
>  #endif
> -#ifdef HAVE_ARCH_HASH_32
> -#if HAVE_ARCH_HASH_32 != 1
> -       pr_info("hash_32() is arch-specific; not compared to generic.");
> -#endif
> -#else
> -       pr_info("hash_32() has no arch implementation to test.");
> -#endif
>  #ifdef HAVE_ARCH_HASH_64
>  #if HAVE_ARCH_HASH_64 != 1
>         pr_info("hash_64() is arch-specific; not compared to generic.");
> diff --git a/tools/include/linux/hash.h b/tools/include/linux/hash.h
> index ad6fa21d977b..38edaa08f862 100644
> --- a/tools/include/linux/hash.h
> +++ b/tools/include/linux/hash.h
> @@ -62,10 +62,7 @@ static inline u32 __hash_32_generic(u32 val)
>         return val * GOLDEN_RATIO_32;
>  }
>
> -#ifndef HAVE_ARCH_HASH_32
> -#define hash_32 hash_32_generic
> -#endif
> -static inline u32 hash_32_generic(u32 val, unsigned int bits)
> +static inline u32 hash_32(u32 val, unsigned int bits)
>  {
>         /* High bits are more random, so use them. */
>         return __hash_32(val) >> (32 - bits);
> --
> 2.33.0
>